TL;DR
Argonne National Laboratory proof-of-concept that applies Microsoft SEAL’s CKKS FHE to a small fully-connected MNIST classifier, demonstrating feasibility while documenting roughly five orders of magnitude runtime overhead and large ciphertext blow-up [Abstract][§4]. The motivating application is privacy-preserving third-party analysis of proprietary reactor structural-health and nondestructive evaluation (NDE) data [§1].
Problem and motivation
Third-party / expert analysis is cost-effective for specialised reactor structural health monitoring and nondestructive evaluation work, but exposes proprietary data to privacy risk [§1, p. 4]. The authors propose using Fully Homomorphic Encryption so a client can hand encrypted data to a third-party processor that performs ML/AI inference without ever seeing the plaintext, then returns an encrypted result for the client to decrypt [Figure 1, p. 4]. The threat model is implicit: a client/server split where the server performs operations on ciphertext only; no explicit honest-but-curious / malicious distinction is stated [§1].
Key contributions
- Apply Microsoft SEAL’s CKKS implementation to a small FCNN performing MNIST digit classification end-to-end on ciphertext [Abstract][§3].
- Document the storage blow-up of CKKS encryption: 10 standardised MNIST images grow from 68.4 KB plaintext to 2.45 GB at poly-modulus degree 8192 and 7.73 GB at 16384 [Figure 2, p. 6].
- Use a degree-2 polynomial approximation of ReLU (fit by L2-minimisation over [-20, +30]) to work around CKKS’s lack of comparison / division operations [§2, p. 6-7].
- Provide a plaintext baseline comparison across FCNN, Approximated FCNN, PCA+NN, and CNN, plus a timing comparison between plaintext and HE FCNN inference [Figure 6, Figure 7, p. 10].
- Flag PCA as a feature-reduction strategy and signal future intent to apply FHE to thermal-imaging defect detection [§5].
FHE setup
- Scheme(s): CKKS (Cheon-Kim-Kim-Song, approximate arithmetic) [§1, p. 4-5].
- Library / implementation: Microsoft SEAL [§1, p. 4].
- Parameters: poly-modulus degree 8192 (4 coefficients, supports ~2 multiplications) or 16384 (6 coefficients, supports ~4 multiplications) [§2, p. 6]. Security level: Not reported. Scale: Not reported.
- Bootstrapping used: Not reported (the paper instead limits multiplicative depth via the chosen poly-modulus degree and coefficient set) [§2].
- Packing / encoding strategy: Not reported.
ML setup
- Task: Image classification inference on encrypted inputs (MNIST 0-9) [§3, §4].
- Model architecture: Small Fully-Connected Neural Network with one hidden layer of 100 nodes using ReLU, then an output layer of 10 nodes using SoftMax [Figure 4, p. 8]. A CNN variant (two conv layers, one max-pool, two FC layers) was trained in plaintext only [Figure 5, p. 8]. PCA was implemented in plaintext only [§3, p. 9]. Layer count convention here: weight-bearing layers only (hidden FC + output FC = 2).
- Activation handling: ReLU replaced by a degree-2 polynomial p(x) = 2.593 + 0.4752x + 0.0173x^2, fit by minimising the squared L2 difference over [-20, +30] [§2, Eq. 1-3, p. 7]. SoftMax is not approximated because the authors argue argmax of SoftMax equals argmax of its pre-image [§3, p. 8].
- Operates on: Plaintext model (weights) + encrypted data — “training of the models on the plaintext data before transferring the weights to the HE implementation” [§1, p. 5].
- Training vs inference: Training is plaintext; only inference runs under encryption [§1, p. 5].
Datasets
| Dataset | Task | Size (train/test) | Modality | Notes |
|---|---|---|---|---|
| MNIST Handwritten | 10-class digit classification | 1000 train / 100 test (subset used for evaluation) | 28x28 black-and-white images | Used as proof-of-concept; “standardized” images [§4, p. 10] |
Pipeline diagram
flowchart LR
subgraph Client
A[Load 28x28 image] --> B[Encrypt CKKS]
end
subgraph Server
C[Plaintext weights] --> D[FC1 100 nodes]
B -. ciphertext .-> D
D --> E[Poly-ReLU deg 2]
E --> F[FC2 10 nodes]
end
subgraph ClientDec [Client]
G[Decrypt CKKS] --> H[Argmax label]
end
F -. ciphertext .-> G
Pipeline steps (text)
- Client loads a standardised 28x28 MNIST image [§4, p. 10].
- Client encrypts the input under CKKS using Microsoft SEAL at poly-modulus degree 8192 or 16384 [§2, p. 6].
- Client transmits the ciphertext to the third-party server [Figure 1, p. 4].
- Server applies the first fully-connected layer (100 nodes) using plaintext weights against the ciphertext [§3, p. 8].
- Server applies the degree-2 polynomial approximation of ReLU element-wise; rescales / relinearises after each multiplication [§2, §3, p. 6-8].
- Server applies the output fully-connected layer (10 nodes); SoftMax is not applied under encryption because argmax is preserved by the pre-image [§3, p. 8].
- Server returns the encrypted logit vector to the client [Figure 1, p. 4].
- Client decrypts and takes argmax to obtain the predicted digit [§3, p. 8].
Architecture diagram
flowchart LR
I([Input 28x28=784]) --> FC1[FC 100]
FC1 --> A1[poly-ReLU deg 2]
A1 --> FC2[FC 10]
FC2 --> O([argmax / SoftMax client-side])
CNN (plaintext only, for context)
flowchart LR
I([Input 28x28]) --> C1[Conv]
C1 --> C2[Conv]
C2 --> P[MaxPool]
P --> FC1[FC]
FC1 --> FC2[FC]
FC2 --> O([Output 10])
CNN layer widths and kernel sizes are not specified in the text [Figure 5, p. 8]. The CNN was not converted to an FHE implementation [§3, p. 8].
Results
| Metric | This paper | Baseline | Hardware |
|---|---|---|---|
| Plaintext FCNN accuracy (MNIST) | 83.8% | - | Not reported [Figure 6, p. 10] |
| Approximated FCNN accuracy (poly-ReLU, plaintext eval) | 76.1% (-7.7 pts vs FCNN) | 83.8% (FCNN) | Not reported [Figure 6, p. 10] |
| PCA + FCNN accuracy | 85.9% | 83.8% (FCNN) | Not reported [Figure 6, p. 10] |
| CNN accuracy (plaintext only) | 59.1% | - | Not reported [Figure 6, p. 10] |
| Plaintext FCNN inference time | 2.001x10^-3 s per image | - | Not reported [Figure 7, p. 10] |
| HE (CKKS) FCNN inference time | 295.3 s per image (~5 min) | 2.001x10^-3 s plaintext | Not reported [Figure 7, p. 10] |
| HE slowdown | ~150,000x | 1x plaintext | Not reported [§4, p. 10] |
| Storage blow-up (10 images) | 2.45 GB (N=8192) / 7.73 GB (N=16384) | 68.4 KB plaintext | Ratios 35,818x and 113,011x [Figure 2, p. 6] |
The accuracy reported for the “Approximated Neural Network” in Figure 6 is described in §4 as identical between the plaintext-poly and HE-poly runs (“we receive an identical result between the NN and Approximated NN models” referring to outputs, not accuracy) [§4, p. 10] — the 76.1% number is therefore the accuracy of the network with poly-ReLU, whether evaluated in plaintext or under HE.
Limitations and assumptions
- Storage blow-up is acknowledged as unresolved: “we do not present a solution for this challenge in this paper” [§2, p. 6].
- The ~150,000x runtime slowdown for HE inference is acknowledged and flagged for future work [§4, p. 10].
- CNN was trained only in plaintext (59.1% accuracy is “mediocre”); not converted to FHE [§3, p. 8].
- PCA was implemented only in plaintext; HE version not yet built [§3, p. 9].
- Evaluation uses only 1000 training and 100 test images — a small subset of MNIST [§4, p. 10].
- Degree-2 polynomial approximation costs 7.7 accuracy points; higher orders are flagged as future work [§4, p. 10].
- Hardware (CPU/GPU, RAM, cores, threads) used for the 295.3 s timing is not reported — making the figure hard to compare against other CKKS work [Figure 7, p. 10].
- Threat model is not formally stated (no honest-but-curious vs malicious distinction) [§1].
- Communication cost (multi-GB ciphertexts per batch of 10 images) is noted but not analysed end-to-end [§2, p. 6].
- Multiplicative-depth budget is fixed by poly-modulus choice; the paper does not state how many multiplications the deployed FCNN actually consumes [§2, p. 6].
Related work it compares against
The paper cites but does not run head-to-head benchmarks against:
- CryptoNets (Dowlin et al., 2016) [ref 12].
- Lee et al. RNS-CKKS comparison optimisation (2022) [ref 13].
- Lee et al. “Precise Approximation of CNNs for Homomorphically Encrypted Data” (2021) [ref 14].
- Lee et al. “Privacy-Preserving Machine Learning with FHE for Deep Neural Network” (2022) [ref 15].
- Chiang, “On Polynomial Approximation of Activation Function” (2022) [ref 11].
No quantitative comparison table against these systems is provided [§4-5].
Code and artifacts
Not released. No repository URL is given in the report [§1-5].
Extra diagrams (optional)
Threat model
flowchart LR
C[Client] -- enc(image) --> S[Third-party Server]
S -- enc(logits) --> C
W[Plaintext weights] --> S
The server holds plaintext model weights; only the input/output are encrypted. No explicit adversary model is stated in the text [§1, p. 4-5].
Activation approximation
ReLU is approximated over [-20, +30] by p(x) = 2.593 + 0.4752x + 0.0173x^2 (degree 2), fit by minimising the integrated squared error [§2, Eq. 1-3, p. 7]. See Figure 3 in paper for the visual comparison [Figure 3, p. 7].
Open questions
- What hardware (CPU model, cores, RAM, threads) was used for the 295.3 s figure? Not reported.
- What security level corresponds to the chosen poly-modulus degrees (8192, 16384) in this SEAL configuration? Not reported.
- What is the multiplicative depth actually consumed by the deployed FCNN, and how was the coefficient chain sized? Not reported.
- Is the 76.1% “Approximated NN” accuracy the HE-evaluated accuracy or the plaintext evaluation of the poly-ReLU network? §4 implies outputs match between the two, but this is not stated explicitly.
- What batching / packing (if any) is used inside SEAL? Not reported.
- Is the 295.3 s per-image cost amortised (batched) or strictly single-sample? Not reported.